When do many agents become one? Holonic composition

Understand what a holon is, how a Belief–Desire–Intention agent becomes a part-and-whole in a hierarchy, and the exact conditions a source paper proves are necessary and sufficient for a group of sub-agents to have a single well-defined collective intention.

Track

Research

Module

5 · Research case studies

Prerequisites

Cognitive architectures and BDI · What is an autonomous agent

Say three agents coordinate to process a contract: one reads clauses, one checks compliance, one drafts a reply. We casually say "the team intends to approve the contract." But does the team really hold an intention, or are there just three separate intentions that happen to point the same way today and could diverge tomorrow? The source paper argues this is not a wording quibble: whether a genuine collective intention exists determines whether you can reason about the group as a single unit at all.

The paper borrows two older ideas and fuses them. From the planning theory of intention: an intention is not just a goal but a commitment that resists reconsideration and structures future planning. From holonic systems theory: a holon is an entity that is a whole with respect to its parts and a part with respect to a larger whole — the same agent is both, depending on where you stand. Combining them gives a holonic BDI agent, and the central question becomes: when do the intentions of the sub-holons add up to a real intention of the holon?

A BDI agent is modelled as beliefs, desires, and a distinguished set of intentions that must be internally consistent (they do not contradict each other or the agent's beliefs) and means–end coherent (every intention has a plan the agent believes will achieve it). A holon wraps one or more such agents (or smaller holons — the definition is recursive) together with a composition operator, a shared information space called the substrate, and a learnable neuro-symbolic interface. The base case of the recursion is a single BDI agent.

The neuro-symbolic angle matters because modern agents put a neural model (an LLM, a perception net) underneath symbolic reasoning. The paper refines Kautz's taxonomy of neuro-symbolic systems. Ordinary "Type 2" means a symbolic core that calls neural modules as subroutines, with the symbolic–neural interface fixed at design time. The paper introduces a learnable variant — it calls it Topology 2-prime — in which the interface itself is learned: a function decides which neural module to invoke and with what argument, as a function of past experience (a trace memory). Type 2 is recovered as the special case where that interface is frozen and the memory is unused.

Holon

A Janus-faced entity: a whole made of sub-holons, and simultaneously a part of a larger holon. The base case is a single BDI agent.

Substrate

A shared information space (think blackboard, tuple-space, shared event log, or distributed cache) visible to all members of the holon. The paper keeps it abstract and depends only on its structural properties: write-visibility, ordering, update atomicity.

Learnable interface (Topology 2-prime)

The symbolic–neural boundary promoted from a fixed design-time coupling to an object the agent learns: which neural module to call, and with what input, chosen from the record of past invocations and outcomes.

Well-defined holon intention

A collective intention set that is internally consistent, means–end coherent, and uniquely determined by the sub-holon intentions plus the substrate dynamics.

The paper's first main result (it labels it T1) states that composing two holons yields a well-defined holon-level intention set if and only if three conditions hold together. The "if and only if" is the point: each condition is not optional garnish but genuinely required, and the paper argues each one individually.

C1 — Meshing

The sub-holon intention sets must mesh: there is a joint plan extending both agents' plans such that their combined intentions and beliefs are consistent, and no sub-goal of one agent's plan defeats a sub-goal of the other's. This is the paper's formalisation of a condition from the theory of shared cooperative activity.

C2 — Mutual responsiveness

Each sub-holon can observe the other's changing intentions through the substrate within bounded latency. Without a write–read channel between them, one agent cannot track the other, and the "joint" plan admits executions that silently drift apart.

C3 — Substrate compatibility

When the two substrates are joined, shared identifiers unify in a well-defined, confluent way — so the merged information space is unambiguous and the composition yields a unique result rather than depending on the order updates happened to arrive.

Status: this is PROVEN as a theorem (the preprint gives a proof sketch in both directions). It is a mathematical characterisation, not a measured result. Its practical reading, which the paper states, is that an architecture failing any of C1–C3 can still look coherent under a lucky sequence of events while admitting observation-dependent inconsistencies under an unlucky one — a failure mode that is hard to catch in testing precisely because it is sequence-dependent.

Consider two document-processing agents merged into one reviewing holon. Agent X intends to approve contracts under a fast path; agent Y intends to hold anything touching a regulated clause. Walk the three conditions:

synthetic — checking C1–C3 on a two-agent holon
C1 meshing:        is there a joint plan where "fast-approve" and
                   "hold regulated" do not defeat each other?
                   -> yes, IF the fast path excludes regulated clauses.
                   -> NO, if both can fire on the same document.

C2 responsiveness: can X see when Y places a hold, within bounded time,
                   via the shared substrate?  -> needs a write-read channel.

C3 substrate:      do "document #4812" in X's space and in Y's space
                   unify to the same object, deterministically?  -> needs
                   a confluent identifier-unification rule.

If C1 fails, there is no collective intention to speak of — the holon is two agents that will contradict each other on the first overlapping document. If C1 holds but C2 fails, they agree in principle but cannot keep each other informed, so their executions diverge. If C1 and C2 hold but C3 fails, "the document" means different things to each, and the merged intention set is not even uniquely defined. Only all three together give you a holon you can reason about as one.

Full meshing (C1) is a strong demand. In practice, agents of independent origin frequently have local conflicts that do not actually preclude working together under sensible arbitration. The paper handles this with a second result (it labels it T6, Arbitrated Composition): introduce a tribunal — a function that takes two possibly-conflicting intention sets plus the shared beliefs and returns a consistent resolution, dropping or substituting the minimum number of intentions needed to restore consistency.

The tribunal is required to be conservative (if the intentions already mesh, it changes nothing), consistent (its output never contradicts the shared beliefs), and maximal (it removes no intention unless removal is necessary — a Razor against over-pruning). The paper proves that an arbitrated composition still yields a well-defined holon intention set under C2 and C3, and that the clean T1 case is recovered exactly in the limit where everything meshes. The concrete arbitration policy — priority ordering, voting, learned arbitration, contract-based reasoning — is left open; the formalism only fixes the three properties the policy must satisfy.

For an engineer this is a design-time checklist, not a runtime surprise: a proposed multi-agent architecture either satisfies C1–C3 (or C2–C3 plus a well-behaved tribunal) and has well-defined collective intentions, or it does not and the failure mode is now nameable. It also connects to security: a tribunal that is not maximal silently discards legitimate intentions, and one that is not consistent can wave through conflicting ones — both are the kind of authorization gap that agent-collusion and privilege-escalation failures exploit.

Stated honestly, at the level the preprint actually establishes:

• ▸PROVEN (deductively, with proof sketches): the Composition Theorem (C1–C3 are necessary and sufficient for a well-defined holon intention) and the Arbitrated Composition result (a conforming tribunal preserves well-definedness and recovers the meshing case as a limit).
• ▸ASSUMPTION / scope: the results assume finite intention sets and depend only on abstract structural properties of the substrate (write-visibility, ordering, atomicity), not on any particular implementation.
• ▸NOT YET MEASURED: these are mathematical characterisations. The paper does not claim empirical validation of them here; an interventional experiment programme is explicitly deferred (see case lesson 3).
• ▸OPEN: the paper leaves the concrete tribunal design, adversarial sub-holons, and non-stationary conditions to future work.

Expected direction: Check C1 meshing (a joint plan exists where neither agent's sub-goals defeat the other's), C2 mutual responsiveness (each can observe the other's intention changes through the substrate within bounded latency), and C3 substrate compatibility (shared identifiers unify confluently). If C1 fails there is no collective intention at all — the intentions cannot even co-exist in one consistent joint plan. C2 failure gives an intention that drifts (they agree but cannot stay informed); C3 failure gives one that is ambiguous / not uniquely defined.

What is a holon, and what is its base case here?

An entity that is simultaneously a whole (made of sub-holons) and a part (of a larger holon). The recursive base case is a single BDI agent.

What does the Composition Theorem claim, and how strong is the claim?

That three conditions — meshing, mutual responsiveness, substrate compatibility — are together necessary and sufficient for sub-holon intentions to form a well-defined collective intention. It is proven deductively (proof sketch), not measured.

What problem does the tribunal solve, and what three properties must it satisfy?

It handles partial meshing (local conflicts that do not preclude cooperation) by resolving conflicting intention sets. It must be conservative, consistent, and maximal — removing the minimum number of intentions needed to restore consistency.

How does "Topology 2-prime" differ from ordinary Kautz Type 2?

Type 2 has a fixed, design-time symbolic–neural interface. Topology 2-prime makes that interface learnable — a function, informed by a trace memory, decides which neural module to invoke and with what argument. Type 2 is the frozen-interface special case.